What Data for This Category Looks Like
Ledger data for petroleum refining comes from DCS/APC control systems, LIMS systems, tank farm metering systems, crude oil evaluation databases, equipment management systems, and energy management systems. Data update frequencies include three categories: real-time, shift-based, and daily. Daily documents include operating procedures, production plans, material balance sheets, distillate export quality reports, interlock ledgers, shift handover records, and change management documents. Fields cover raw material property parameters, unit operation parameters, material flow data, and quality indicators. Units include temperature, pressure, flow rate, component proportion, and similar metrics.
Constraints on Multi-turn Dialogue and Prompt Engineering
Multi-source heterogeneous ledger data with inconsistent update frequencies requires clear specification of data time ranges and source systems during multi-turn dialogue, to avoid confusion across systems. Differences in fields across multiple document types require prompts to preset parsing rules for different documents. For example, extract interlock trigger conditions from interlock ledgers, and extract shift information from shift handover records. High-frequency real-time data updates require dialogue contexts to automatically filter expired data, ensuring the latest DCS and tank farm metering data is used. Logistics and energy coupling between refining units requires multi-turn dialogue to associate ledger data from upstream and downstream units, avoiding isolated analysis of single-unit operations.
Configuration Settings
The following configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Context window size | 8000–12000 characters | Adapts to the total data length of multi-source refinery ledgers, avoiding truncation of key process parameters |
messages field format | Arranged in the order of system prompt + historical dialogue + current user question | Aligns with multi-turn dialogue context transfer logic, ensuring historical query results are correctly associated |
variables variable binding | Bind preset variables such as crude oil evaluation ID, unit number, shift identifier | Quickly locates dedicated ledger data for corresponding batches, reducing manual input errors |
chatId association rule | Bind the session ID of a single shift or single unit | Isolates ledger data from different production units, avoiding data confusion across units |
| Maximum dialogue turns | 5 turns | Controls total context length, preventing response delays caused by excessive historical data |
Three Common Configuration Errors
- Symptom: The AI dialogue node cannot retrieve the initial ledger query request from the user, and can only receive intermediate parameters passed by the judgment component. Cause: The initial user question is not bound via
variablesin the workflow orchestration, resulting in missing original query instructions in the dialogue context. - Symptom: Ledger results returned by the dialogue lack refinery-specific fields such as distillate export dry point and tank farm liquid level, or have incorrect field units. Cause: The prompt does not clearly specify the required refinery ledger fields and corresponding units, causing the model to parse generic document formats.
- Symptom: Multi-turn dialogue responses for locally deployed models are slow, with a single query taking more than 600 seconds. Cause: The context window size is not limited, and too much multi-unit, multi-time-period ledger data is called, exceeding the model's character limit.
How to Confirm Successful Configuration
- Initiate a multi-turn dialogue including queries about crude oil properties and unit load, check whether
chatIdis bound to the session identifier of the corresponding production unit to confirm data isolation is effective. - Check whether the preset variables bound by
variablescorrectly associate the current query's crude oil evaluation ID or unit number, verifying the accuracy of data positioning. - Initiate consecutive queries within 5 turns, check whether the response time meets expectations, confirming that the context window is not excessively occupied.
- Verify that the returned ledger data fields include refinery-specific content such as distillate export components and tank farm liquid level, confirming that the prompt parsing rules are effective.
The values provided are common starting points and should be measured against the reader's own samples.